Re: Bug#647522: non-deterministic compression results with gzip -n9

[Cyril Brulebois]

Paul Eggert <address@hidden> (08/02/2012):
> Thanks very much for the patch.  But can someone who's looked into it
> please explain why 'window' needs to be zeroed out?  This will save me
> time in reviewing the patch.  Thanks.

Welcome. Here's a slightly more detailed analysis. You can find attached
the sample files I used (still from the ppl source package). I managed
to track it down to deflate() (what a surprise!), so I decided to check
the matches. Printing match_length only led me to a single difference,
while printing both match_length and match_start (which according to a
comment is set by longest_match()), I got two differences.

To produce the below-quoted diff, using the attached patch, I did:
| tar xfz ppl.tar.gz
| gzip -9f README CREDITS > match-2
| tar xfz ppl.tar.gz
| gzip -9f CREDITS > match-1
| diff -u match-2 match-1

Of course there's an extra file showing up in the diff, but skipping it,
here's the diff:
|  match: 46 @ 19459
|  match: 8 @ 15659
|  match: 8 @ 15659
| -match: 7 @ 17193
| -match: 6 @ 17193
| +match: 6 @ 19510
| +match: 6 @ 19510
| [ no more matches ]

Interestingly, the size of the CREDITS file is 19579.

Mraw,
KiBi.

0001-Pin-point-dirty-window-bug.patch
Description: Text Data

ppl.tar.gz
Description: Binary data

signature.asc
Description: Digital signature